Open apmoore1 opened 4 years ago
The way I believe Peng et al. 2019 did that task was first extract the targets and their associated sentiment label. At the same time they extract the opinion words. With both a list of opinion words and targets with their sentiment they performed a many to many matching task to find the relevant opinion words for the relevant targets.
Relevant problem of linking targets to their sentiment is co-reference resolution of which I think this is kind of similar to how Peng et al. 2019 performed the task, relevant co-reference paper: End-to-end Neural Coreference Resolution
From issue #3 it would be useful to create a schema that can do what Peng et al. 2019 did which was link the targets and their sentiment to the opinion words. An idea for this would be to solve the opinion word problem as a relation extraction type of task e.g. predict the start and end index of the opinion word given the target and its sentiment or just the target.